Generates images with transparent backgrounds and edits photos from up to ten reference shots

You type a prompt and get a 2K image, or you hand it a photo and say what to change; both jobs run through the same model. The new part is transparency. Ask for a transparent (RGBA) image, meaning a file with a real alpha channel, and the result is a clean cutout that drops straight onto a slide or a product page, not a fake checkerboard. It also works the other way and lifts a subject out of an existing photo.

For editing, you can attach up to ten reference images at once. Six portraits become one group photo with each face kept intact; a product shot plus a background plate become a composited scene. Local edits can be pointed at with a circle drawn on the image, a painted scribble or a separate mask, so "change only this" works without prompt gymnastics. Typography and portrait lighting got attention too, so text on signs and packaging comes out readable.

The image generator itself is 7B parameters, small for this level of output. It encodes the prompt and the reference images once and reuses them across every denoising step, which is what keeps multi-image edits affordable. Weights are bf16 safetensors, and diffusers, ComfyUI, vLLM-Omni and SGLang all supported the model on release day.

Start with the official Space on Hugging Face: one tab for text-to-image, one for editing with up to ten input images, and a gallery of worked examples you can load with a click. One thing to know: the Space calls Qwen's hosted API rather than loading the downloaded weights, so treat it as a preview of what the model does, not a test of the exact checkpoint. First test: upload two photos of different people and ask for both in one scene, then check whether each face survives. Then ask for a sticker on a transparent background and open the PNG in an editor to confirm the background is really empty. To run it yourself, the diffusers pipeline is a ten-line script; in bf16 the generator and its text encoder together want well over 24 GB, so on a single consumer card use the documented CPU offload option or the ComfyUI day-one workflows.

The weights ship under the Qwen Research License, which allows research and evaluation only; commercial use needs a separate license from Qwen. That is a step back from the earlier Qwen-Image releases, which were Apache 2.0, so check before building a product on it.

• Make stickers, icons and logo marks that arrive already cut out, ready to layer.
• Combine several people's photos into one group shot without a photoshoot.
• Drop a product into new backgrounds while keeping its shape and label intact.
• Circle the one thing you want changed in a photo and leave everything else alone.
• Mock up posters and packaging where the text is actually legible.

Try the demo

View model page

Read this issue on neuralexpedition.com →